DYNAMIC MODELS OF ASSET RETURNS AND MORTGAGE DEFAULT. Xi Chen

Transcription

1 DYNAMIC MODELS OF ASSET RETURNS AND MORTGAGE DEFAULT Xi Chen A dissertation submitted to the faculty of the University of North Carolina at Chapel Hill in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the Department of Statistics and Operations Research in the University of North Carolina at Chapel Hill Chapel Hill 2017 Approved by: Eric Ghysels Chuanshu Ji Jonathan Hill Vidyadhar Kulkarni Vladas Pipiras

2 c 2017 Xi Chen ALL RIGHTS RESERVED ii

3 ABSTRACT XI CHEN: Dynamic Models of Asset Returns and Mortgage Default (Under the direction of Eric Ghysels and Chuanshu Ji) This dissertation consists of three chapters. The first chapter builds a new series of dynamic copula models and studies the influence of macro variables on the dependence between assets. The second chapter develops a dynamic logistics regression model and investigates how systematic risk affects mortgage default. The third chapter uses the frailty model developed in chapter 2 to explore spatial dependence between commercial and residential mortgage risk. In all three chapters, we extend the generalized autoregressive score (GAS) models proposed in Creal, Koopman and Lucas (2013a). In the first chapter, we propose a series of dynamic copula models with a shortand long run component specification, inspired by the mixed data sampling (MIDAS) component structure applied to univariate GARCH models in Engle, Ghysels and Sohn (2013) and multivariate GARCH models in Colacito, Engle and Ghysels (2011). In particular, we extend the framework of MIDAS to dynamic copulas. In the framework of GAS models, we combine macro variables of low frequency with asset returns of high frequency, and investigate the influence of low frequency macro variables on the dependence between asset returns. Our data consists of stock portfolios and a bond. We assess the new class of models with these data and find that an extra component enhances the model with more volatility. Moreover, the macro variables with MIDAS work as a proxy for the market condition, and allow that the macro environment affects how dependence parameter reacts to innovations. With these two flexibilities, the model performance is consistently improved through our empirical applications. In the second chapter, we design a new dynamic logistic regression model to track iii

4 systematic risk of mortgages. Specifically, we match default rates in multiple dimensions by extending the GAS models. Our data consists of commercial mortgages in the U.S. retail market from 1997 to An empirical analysis of these data suggests the influence of origination month and the originator preference on default rates. To model the effects of these variables, we group mortgages by these two variables and allow latent factors to vary by groups. Compared with GAS models using a single factor, our multi-factor models feature improved empirical fits. To the best of our knowledge, this is the first attempt that uses observation-driven models to predict mortgage defaults. We show that the new class of models has better tractability compared with parameter-driven models. For instance, although our dataset has more than two million records, and our most complex model incorporates up to 15 frailty factors, the estimation process only takes two minutes using a standard desktop computer. In the third chapter, we use the frailty model developed in chapter 2 to explore spatial dependence between commercial and residential mortgage risk. Our dataset contains 1.6 million records of commercial mortgages and 140 million records of residential mortgages in the U.S. market. The time range of these records is between January 1999 and March Our empirical analysis demonstrates strong spatial dependence between commercial defaults and residential default in multiple respects. First, we apply Granger causality tests to the empirical default rates of commercial mortgages and residential mortgages in 10 main MSA areas, and the test results in 9 areas reveal a significant lead and lag relationship of the two mortgage markets. Second, we test the causal relation among the frailty factors that explain systematic risk of commercial mortgage and residential mortgage, and provide strong evidence on the close correlations between the residential and commercial mortgage markets. Last but not least, we show that residential PD is a good explanatory variable in predicting default of commercial mortgages in adjacent area, and this prediction power also implies that local residential market drives the commercial market. To the best of our knowledge, this is the first paper exploring iv

5 the spatial dependence between commercial mortgage default and residential mortgage default. v

6 ACKNOWLEDGEMENTS First and foremost, I would like to offer my deepest gratitude to my advisors, Professors Eric Ghysels and Chuanshu Ji. It is my fortune to have them as my academic advisors. Their ingenious thoughts and deep insight have inspired me so much and their knowledge of finance and statistics is a huge resource for my research. They are always on my back and support me whenever needed. Besides research, they are also my personal advisors. They cares about my career and always gives me advice, from which I will benefit for the rest of my life. In a word, I learnt so much from Professors Eric Ghysels and Chuanshu Ji, and without them I would not have such a wonderful experience as a Tar Heel. I would also like to show my deep appreciation for other committee members: Professor Jonathan Hill, Professor Vidyadhar Kulkarni, and Professor Vladas Pipiras. They provide invaluable helps in my graduate study and advice for my dissertation. Professor Jonathan Hill read my dissertation carefully and provided numerous useful comments and feedbacks. I learnt stochastic models in operations research and market dynamics from Professor Kulkarni. Professor Pipiras taught me advanced probability. I am extremely grateful to them Lastly and most importantly, I would like to thank my parents and wife. They always love, believe and support me whatever happens and sacrifice a lot for me. I hope to make them proud of me for my accomplishment. I dedicate this dissertation to them. vi

7 TABLE OF CONTENTS LIST OF FIGURES... ix LIST OF TABLES... x 1 Component Dynamic Copula Models with MIDAS Introduction Background Model Formulation Notation and Preliminaries A New Class of Component Dynamic Copula Models Estimation Empirical Application Data and Variables Results Conclusions Frailty Models for Commercial Mortgages Introduction Literature Review Model Formulation Empirical Applications Data and Variables Estimation Results vii

8 2.5 Conclusions Commercial and Residential Mortgage Defaults: Spatial Dependence with Frailty Introduction Background Model Formulation Empirical Applications Data Variables for Commercial Mortgages Variables for Residential Mortgages Estimation Results Conclusion BIBLIOGRAPHY viii

9 LIST OF FIGURES 1.1 Industrial Production and Realized Correlations The Quarterly Realized Correlations of the Stock and Bond The Quarterly Realized Correlation and Implied Correlations The Monthly Realized Correlations and the Implied Correlations The Implied Correlations of GAS-DCC and GAS-DCC-MIDAS Models The Implied Correlations of GAS and GAS-DCC Models The Implied Correlations of GAS-DCC and GAS-ADD Model Empirical Default Rates (PD) by Mortgage Age Default Rates by Mortgage Age of Static I and Static II Default Rates by Exposure Month of Static I and Static II Default Rates by Exposure Month of Static I and Static III Default Rates by Exposure Month of Dynamic I Default Rates by Origination Month of Dynamic I Default Rates by Origination Month of Dynamic II Default Rates by Originator Group of Dynamic II Default Rates by Originator Group and Mortgage Age of Dynamic II Default Rates by Originator Group of Dynamic II and Dynamic III Default Rates by Originator Group and Mortgage Age of Dynamic II and Dynamic III Empirical Default Rates of Commercial and Residential Mortgages in the Top 10 MSA Areas Frailty Factors of Commercial and Residential Mortgages in the Top 10 MSA Areas ix

10 LIST OF TABLES 1.1 Estimation Results for Consumer Goods Estimation Results for Manufacturing Estimation Results for Health Estimation Results for HiTec Estimation Results for Other Components of Static and Dynamic Models The Grouping Criterion for Originator Frailty Estimates for Static and Dynamic Models The Summary Statistics of Commercial Mortgages in the Top 10 MSA Areas The Summary Statistics of Residential Mortgages in the Top 10 MSA Areas Components of Static and Dynamic Models for Commercial Mortgages Components of Static and Dynamic Models for Residential Mortgages Estimates of the Static and Dynamic Models for Commercial Mortgages Estimates of Static and Dynamic Models in the Los Angeles- Long Beach-Glendale Area Estimates of Static and Dynamic Models in the New York- Jersey City-White Plains Area Estimates of Static and Dynamic Models in the Houston- The Woodlands-Sugar Land Area Estimates of Static and Dynamic Models in the Atlanta- Sandy Springs-Roswell Area Estimates of Static and Dynamic Models in the Phoenix- Mesa-Scottsdale Area Estimates of Static and Dynamic Models in the Dallas- Plano-Irving Area Estimates of Static and Dynamic Models in the Riverside- San Bernardino-Ontario Area Estimates of Static and Dynamic Models in the Chicago- Naperville-Arlington Heights Area x

11 3.14 Estimates of Static and Dynamic Models in the Anaheim- Santa Ana-Irvine Area Estimates of Static and Dynamic Models in the Washington- Arlington-Alexandria Area Granger Causality Tests between Commercial PD and Residential PD in Main MSA Areas Granger Causality Tests between Commercial Frailty and Residential Frailty in Main MSA Areas xi

12 CHAPTER 1: COMPONENT DYNAMIC COPULA MODELS WITH MIDAS 1.1 Introduction Measuring temporal dependence between financial assets is a key ingredient to risk hedging, asset pricing, portfolio choices, to name only a few. For example, hedging ratios dynamically adjust to the varying dependence between financial assets. Likewise, the pricing of structured products such as CDO s critically relies on the dependence between the underlying financial assets. To model temporal dependence between asset returns, two main methods have been developed in the literature. One is multivariate GARCH models and the other one is copula-based models. This chapter focuses on the latter. Copula-based models allow researchers to model the marginal distribution and dependence structure separately. This property provides a flexible framework to model multivariate time series and recently increasing attention has been devoted to conditional copulas when modeling dynamic dependence of financial assets. Patton (2006) provided the theoretical foundation for conditional copulas. He used a combination of GARCH models and copulas to model Deutsche Mark and Yen jointly. The dependence parameters of the copulas are driven by autoregressive processes. Guegan and Zhang (2010) compared two dynamic copula models and proposed statistical tests based on conditional copulas. Fengler and Okhrin (2012) utilized realized variance to model the dependence between daily stock returns, and the dynamic of the copula is driven by a HAR (Corsi, 2009) process - which is a MIDAS specification with step functions. To parameterize dynamic copulas in non- Gaussian settings, Creal, Koopman and Lucas (2013a) proposed to use the scores of log likelihood functions as the innovation term, and named the models as Generalized 1

13 Autoregressive Score Models (GAS). In this chapter, we propose a series of dynamic copula models with a short- and long run component specification, inspired by the mixed data sampling (MIDAS) component structure applied to univariate GARCH models in Engle, Ghysels and Sohn (2013) and multivariate GARCH models in Colacito, Engle and Ghysels (2011). Hence, the purpose of this chapter is to extend the framework of MIDAS to dynamic copulas. In the framework of GAS models, we combine macro variables of low frequency with asset returns of high frequency, and investigate the influence of low frequency macro variables on the dependence between asset returns. Our data consists of stock portfolios and a bond. The stock data are the daily returns on five industry portfolios. The bond data are the daily returns of a 10-year Treasury bond. We assess the new class of models with these data and find that an extra component enhances the model with more volatility. Moreover, the macro variables with MIDAS work as a proxy for the market condition, and allow that the macro environment affects how dependence parameter reacts to innovations. With these two flexibilities, the model performance are consistently improved through our empirical applications. The rest of this article is organized as follows. In the next section, we review the literature. Section 1.3 states the model formulation. In section 1.4, we describe the details of estimation. Section 1.5 discusses empirical applications. The last section provides concluding remarks. 1.2 Background To motivate the theoretical framework, it is useful to review the literature on copulas and related areas. We first introduce the theoretical foundation of static and dynamic copulas. Next, we focus on various parameterization methods for dynamic copulas. At last, we briefly cover component models in the context of volatility forecasting. 2

14 We start with the introduction of static copulas. In particular, consider a multivariate random variable Y = [Y 1,..., Y n ]. Let F be the joint cumulative distribution function (CDF) for Y, and F i be the CDF for Y i. By Sklar s theorem (Sklar 1959), there exists a copula function C( ) : [0, 1] n [0, 1], mapping the marginal distributions of Y i to the joint distribution through: F (y) = C(F 1 (y 1 ),..., F n (y n ) ρ) (1.2.1) where ρ is the dependence parameter of interest. Accordingly, the joint probability density function (P DF ) can be represented as the product of copula density c( ) and marginal PDF f i : n f(y) = c(u 1,..., U n ρ) f i (y i ) (1.2.2) where U i = F i (y i ). Note that it is usually assumed that c( ) and f i share no common parameters. In the case of (1.2.2), ρ does not appear in f i. i=1 If we take log of both sides of equation (1.2.2), the product of c(u 1,..., U n ρ) and f i (y i ) transforms to the sum of log(c(u 1,..., U n ρ)) and log(f i (y i )). This transformation motivates a two-stage estimation: The first stage estimates the parameters of marginal distributions with the likelihoods only involving log(f i (y i )), and the second stage estimates the parameters of c( ) with the likelihoods only containing log(c(u 1,..., U n ρ)). This two-stage estimation greatly reduces the computation cost for estimation, because the parameters of f i and c( ) can be estimated in separate optimizations. Comparing with a joint estimation of all parameters, this two-stage estimation may entail some efficiency loss. However, as the numbers of parameters increase with problem sizes, the two-stage estimation may be the only feasible estimation method in practice. While Skalar s theorem motivates the application of static copula models, Patton (2006) further establishes the theoretical foundation of dynamic copula models. Suppose Y t = [Y 1,t,..., Y n,t ] is a multivariate stochastic process and F t 1 is the information set up 3

15 to time t 1. Patton (2006) showed that the conditional CDF F (. F t 1 ) can be decomposed into the conditional marginal CDF F i (. F t 1 ) and conditional copula C(. F t 1 ) as what follows: F (y t F t 1 ) = C(F 1 (y 1,t F t 1 ),..., F n (y n,t F t 1 ) F t 1 ) (1.2.3) n f(y t F t 1 ) = c(u 1,t,..., U n,t ρ t, F t 1 ) f i (y i,t F t 1 ) (1.2.4) where ρ t is a dynamic dependence parameter and changes with time. i=1 In Patton (2006) both Gaussian and non-gaussian copulas are constructed to model the dependence between Deutsche mark and Yen. In the Gaussian case, ρ t has the following dynamic, ρ t = Λ 1 (δ t ) δ t = ω + βδ t 1 + α 10 Φ 1 (U 1,t k )Φ 1 (U 2,t k ) 10 k=1 where Λ 1 ( ) is a link function to make sure that ρ t is between -1 and 1. δ t is the transformed dependence parameter. Φ 1 ( ) is the inverse CDF function of normal distribution and U i,t k = F i (y i,t k F t 1 ). This dynamic is of autoregressive form, and it is driven by a lagged part and an innovation part. For the non-gaussian case, symmetrized Joe- Clayton (SJC) copula is used and Patton (2006) suggested the following dynamic for the tail dependence parameter ρ t : ρ t = Λ 2 (δ t ) δ t = ω + βδ t 1 + α 10 U 1,t k U 2,t k 10 k=1 4

16 Here Λ 2 (x) is a link function to ensure ρ t lies in its domain. Besides, Patton (2006) used an ARMA-GARCH model for the marginal distribution of assets returns. Since the seminal work of Patton (2006), various dynamic copulas have been proposed, and most of them focus on the parameterization of dependence parameters. Heinen and Valdesogo (2009) suggested using DCC framework to model the dependence parameter. Christoffersen et al. (2012) adapted the DCC framework to reduce the computational complexity. While all these preceding models are observation driven, Hafner and Manner (2012) proposed a parameter driven model with a latent stochastic process. Hafner and Reznikova (2010) also developed a semi-parametric approach to model the dependence parameter as a smooth function of time. Structural breaks were used to model the dependence parameter in Dias and Embrechts (2002) and Manner and Candelon (2010). Many of these papers assume that the dynamics of dependence parameters are of autoregressive form that contains a lagged term and an innovation term. Between these two terms, the choice of the innovation term is crucial and depends on the functional forms of copulas. For example, Patton (2006) used cross products and differences in the Gaussian and non-gaussian cases respectively. In the latter case, differences are used because the interpretation of cross products is not clear with non-gaussian distributions. To formulate the dynamics of parameters in general settings, Creal, Koopman and Lucas (2013a) and Harvey (2013) proposed GAS models, which use the scores of log likelihood functions as the innovation term. These researchers assumed δ t -the transformed dynamic parameter- of the following form: δ t = ω + p q A i s t i + B j δ t j i=1 j=1 Since δ t may be a vector, all the terms here are of appropriate dimensions. A i and B j are coefficients of the innovation term and lagged term. It is further assumed that s t is 5

17 the scaled score of the likelihood function, as shown below: s t = S t t t = ln f(y t δ t, F t 1 ; θ) δ t S t = S(t, δ t, F t 1 ; θ) where t is the score of the log likelihood function and S t is a matrix function to scale the score. Several choices for the scaling matrix S t are proposed: It can be the inverse information matrix, the square root of the inverse information matrix, or an identity matrix, as displayed below: S t = I 1 t t 1, I t t 1 = E t 1 [ t t 1] or S t = J 1 t t 1, J t t 1J t t 1 = I 1 t t 1 or S t = I where I is an identity matrix. While GAS models apply to general problems involving time varying parameters, they are of special importance to copula modeling. A large number of copulas are constructed from non-gaussian settings, and it is hard to find an innovation term. GAS models have been applied to dynamic copulas in Oh and Patton (2013), Salvatierra and Patton (2014) and Patton (2012). In this chapter we also use this framework and compare its performance with the model driven by cross products. Component models have also attracted considerable attention in the literature when modeling volatility and correlation of asset returns. In Engle and Lee (1999), they proposed a GARCH model driven by two components and could be seen as a restricted GARCH(2,2) model. In Engle and Rangel (2008), a multiplicative component GARCH model was proposed and they related return volatility to macroeconomics. Similarly, 6

18 Colacito, Engle and Ghysels (2011) developed an additive component models named DCC-MIDAS, separating long term and short term components. In this chapter, we extend the frameworks of GARCH-MIDAS and DCC-MIDAS to dynamic copula modeling. 1.3 Model Formulation The purpose of this section is to introduce a new series of dynamic copula models. In a first subsection, we provide some preliminaries and describes two benchmark models in the literature. The second subsection introduces the structures of new dynamic copula models Notation and Preliminaries To set up models, consider a bivariate stochastic process y t = (y 1,t, y 2,t ). We assume the marginal distribution of y i,t follows the GARCH-MIDAS framework in Engle, Ghysels and Sohn (2013). Specifically, the dynamic of y i,t is as follows: y i,t = µ i + τ i,t g i,t ɛ i,t (1.3.1) (y i,t 1 µ i ) 2 g i,t = (1 α i β i ) + α i + β i g i,t 1 τ i,t (1.3.2) τ i,t = m + θ i K k=1 ψ k (ω i,1, ω i,2 )X t k (1.3.3) where µ i is the constant mean of y i,t. The volatility dynamic for y i,t has two components. The short term component g t assumes an autoregressive form shown in (1.3.2). The long term component τ t is driven by a weighted sum of X t k, and X t k could be external information or derived from y i,j (j t 1). The weight ψ k (ω i,1, ω i,2 ) is determined by parameter ω i,1 and ω i,2. Denote the information set up to time t 1 as F t 1, and we assume ɛ i,t F t 1 F i ( ). For the joint distribution, we assume that ɛ t = (ɛ 1,t, ɛ 2,t ) is generated by a dynamic 7

19 copula, i.e.: F (ɛ 1,t, ɛ 2,t ρ t, θ, F t 1 ) = C(F 1 (ɛ 1,t ), F 2 (ɛ 2,t ) ρ t, ν, F t 1 ) where C(, ) is a bivariate copula and ρ t is the dependence parameter of interest. ν includes static parameters. As will be discussed in the estimation section, we choose t copula for the joint distribution and skewed t distribution for the marginal distribution of ɛ i,t. Note that the dependence parameter ρ t for t copula lies in the interval [ 1, 1]. Regarding the parameterization of ρ t, two benchmark models arise in the literature since we choose t copula for the joint distribution. One is the cross product model used for Gaussian copula in Patton (2006) and t copula in Christoffersen et al. (2012), with the dynamics as below: ρ t = Λ 1 (δ t ) (1.3.4) δ t = ω + βδ t 1 + α 1 Ψ 1 (U 1,t 1 )Ψ 1 (U 2,t 1 ) (1.3.5) Λ 1 (x) = 1 exp(x) 1 + exp(x) (1.3.6) where Ψ 1 ( ) is the inverse CDF of t distribution and U i,t 1 = F i (ɛ i,t 1 ). 1 We call models based on this dynamic as PROD models in short of cross product. Another benchmark model is GAS model as discussed in section 1.2. Assuming one lag period, we have the following dynamics: δ t = ω + βδ t 1 + α 1 s t 1 (1.3.7) where s t 1 is the score of the likelihood function with respect to δ t A New Class of Component Dynamic Copula Models In this subsection, we introduce the new class of dynamic copula models. At first, we propose an additive component model based on GAS model, motivated by the for- 1 The degree of freedom and skewness of this t distribution are the same as the ones in the t copula we specified in the preceding paragraph, i.e.: C(, ρ t, ν, (F). 8

20 mulation of DCC-MIDAS model. We add into equation (1.3.7) a time varying intercept that is the moving average of lagged δ t as below: δ t = ω k k δ t i + βδ t 1 + α 1 s t 1 (1.3.8) i=1 This model is a natural extension of DCC-MIDAS model to dynamic copulas. The idea underlying DCC-MIDAS is extracting two components from the daily correlations: one short term component from the daily innovation and a long term component driven by the moving average of realized correlations. Similar logic applies here with two notable differences. One difference is that we use scores rather than cross products, because scores incorporate more information from the functional form of t-distribution than cross products. The other difference is the construction of long term component. While DCC- MIDAS model uses realized correlation for the dynamic intercept, we use the fitted dependence parameter to simplify computation. Since most of the times, there is no closed form relation between realized correlation and the dependence parameter of the copula. We call this new additive component model as GAS-ADD model. Similarly, we can also take an average of k lagged innovations to form a long term innovation, yielding another component model as follows: δ t = ω + βδ t 1 + α 1 s t 1 + α 2 k k s t i (1.3.9) i=1 This model shares similar properties with DCC model. It could be seen as a specialized DCC model using scores as innovation with parameter restrictions. So we call it GAS-DCC model. GAS-DCC and GAS-ADD models decompose the daily innovations of dependence parameters into long and short term components; moreover, we also want to extract the influence of macro variables on the dependence parameters. Among various ways to 9

21 include macro variables, we choose to follow the GARCH-MIDAS framework in Engle, Ghysels and Sohn (2013), and incorporate macro variables multiplicatively: δ t = ω + βδ t 1 + α 1 s t 1 (1.3.10) τ t K log(τ t ) = m + θ 1 ψ 1( K k (ω 1, ω 2 )X t k + θ 2 1( K ψ k (ω 1,ω 2 )X t k )<0 k=1 k=1 ψ k (ω 1,ω 2 )X t k )>0 k=1 k=1 (1.3.11) K ψ k (ω 1, ω 2 )X t k where X t k is a macroeconomic variable. ψ k (ω 1, ω 2 ) is the weight assigned by MIDAS polynomial to X t k with parameters ω 1 and ω 2. The above formulation has a natural interpretation in GARCH model, since asset returns tend to react differently to news depending on the macroeconomic environment. Now τ t influences ρ t similarly but in a nonlinear way, due to the existence of link function. Note that in Engle, Ghysels and Sohn (2013), the intercept in equation (1.3.2) is specified as 1 α β. However, that specification is based on the assumption of Gaussian distribution that no longer holds here. So we use a separate parameter ω. We call this model GAS-MIDAS thereafter. GAS-MIDAS model lets macro variables influence short term innovations. Alternatively, we can also weight macro variables by the long term innovations in the GAS-DCC model. In particular, we divide α 2 in equation (1.3.9) by macro variables and have the following dynamics: δ t = ω + βδ t 1 + α 1 s t 1 + α 2 kτ t log(τ t ) = m + θ 1 1 K + θ 2 1 K ψ k (ω 1,ω 2 )X t k <0 k=1 k=1 ψ k (ω 1,ω 2 )X t k >0 k=1 k=1 k s t i (1.3.12) i=1 K ψ k (ω 1, ω 2 )X t k (1.3.13) K ψ k (ω 1, ω 2 )X t k The idea here is weighting long term innovation with long term influence of macro variables, and we call this model GAS-DCC-MIDAS. Through equation (1.3.8) to equa- 10

22 tion (1.3.13), we propose three new component models for dynamic copulas. We will compare the performance of these new models with the benchmark models by empirical applications in section Estimation To estimate the parameters of these copula models, we apply the two-stage method discussed in section 1.2. In the first stage, we estimate the univariate GARCH-MIDAS models with quasi maximum likelihood method, and normalize the dependent variables using fitted standard errors. Then we fit a marginal distribution for each of the normalized variables. In the second stage, we fit the copula model with standard maximum likelihood method. This two-stage estimation is generally applied in literature and makes the computation much easier than joint estimation. In the following paragraphs, we discuss the details of this estimation method. There are numerous choices for the marginal distribution, such as normal distribution, the standardized t distribution (Bollerslev (1987)), and the skewed t distribution (Patton (2004)). We use the skewed t distribution for its flexibility. This distribution has two shape parameters controlling its skewness and tail thickness. A skewness parameter, λ ( 1, 1), describes the degree of asymmetry, and a degrees of freedom parameter, ν (2, ), measures the tail thickness. If λ = 0, we have the standardized Student s t distribution. When ν, we have skewed normal distribution. If ν and λ = 0, we recover a standard normal distribution. All these flexibilities make skewed t distribution a good choice to model univariate variables. For further results on this distribution, refer to Hansen (1994) and Jondeau and Rockinger (2006). We choose t copula for the multivariate modeling because of its capability to incorporate various dependence structures. First, it has a degrees of freedom parameter ν c controlling the tail thickness. Second, by varying the dependence parameter ρ, t copula 11

23 can model data of both negative and positive correlation. If ρ equals to one (minus one), we have perfectly positively (negatively) correlated data series. This property is important since the data in empirical applications exhibits both positive and negative correlations. Not all copulas have such flexibility. For example, Clayton and a number of other Archimedean copulas can only model positively or negatively correlated data. For further results on these copulas, refer to Joe (2014) and Nelsen (2007). For the MIDAS component containing macroeconomic variables, we use the same variable for both univariate modeling and multivariate modeling. This consistency ensures the conditional copula a valid one. For discussions on the validity of conditional copulas, see Patton (2006). To select the number of lag periods for the MIDAS components, we follow the profiling method discussed in Engle, Ghysels and Sohn (2013) and Colacito, Engle and Ghysels (2011). For the MIDAS polynomial, we choose a Beta weighting scheme of the following form: ψ k (ω 1, ω 2 ) = (k/k)ω 1 1 (1 k/k) ω 2 1 K (k/k) ω 1 1 (1 k/k) ω 2 1 j=1 Beta weighting scheme offers flexible shape for the MIDAS filters. It can provide both decreasing and increasing schemes. Moreover, it can also offer a hump shaped weighting shape limited to be unimodal. Besides, there are other weighting schemes available. See Ghysels, Sinko and Valkanov (2007) for a further discussion on the choices of weighting schemes. Besides, for the GAS-ADD model, we choose k = 22 for the time varying intercept. For GAS-DCC model, we choose k = 5 for the long term innovation. These lagging periods are picked by the profiling likelihood methods and the clear interpretation of being monthly and weekly averages. 12

24 1.5 Empirical Application Data and Variables In this section, we use the new class of models to investigate the dependence between stocks and bonds. The bond data are the daily returns of a 10-year Treasury bond. The stock data are the daily returns of five industry portfolios compiled by Kenneth French, which could be downloaded from his web page. The five industries include consumer goods, manufacturing, high tech, health and others. The time range of the data is from November 30, 1985 to December 30, 2013, with 7042 observations. Because of the similar patterns across these five industries, we mainly use the pair of manufacturing industry and 10-year treasury bond as an example. If we mention stock, we mean the stock portfolio of manufacturing industry. This applies to all the figure examples in the following paragraphs. We use monthly growth rate of industrial production (IP) in U.S. as the macro variable. 2 For univariate modeling, we compute the quarterly rolling average of the IP rates and apply a MIDAS polynomial with the quarterly average. Specifically, if X t is the variable in the MIDAS polynomial as in equation (1.3.3) and X t is the monthly IP rate, then X t = (X t + X t 1 + X t 2)/3. For each day, we look back for 16 months, i.e.: K = 16 in the MIDAS polynomial. Therefore, there are actually two filters smoothing the macro variables. Similar applications of filters can be found in Engle, Ghysels and Sohn (2013). For multivariate modeling, we use the first order difference of the IP growth rates and take the quarterly average of the difference to smooth the data. Unlike the univariate modeling, we only assume a flat weight for the MIDAS filter. We make these transformations based on empirical investigations. Both models of the raw rates and 2 The IP rates are calculated year over year. 13

25 differences are tested, and the latter one shows a better performance. For the number of lag periods, empirical tests favor a short window of three months. This short window size is also supported by the volatile fluctuation of the realized correlations, which is shown in the the upper panel of Figure 1.1. Clearly, the correlations have many spikes, even if they are calculated on a quarterly basis. Meanwhile, the IP growth rates in the lower panel of Figure 1.1 change relatively slow. It is hard to relate the change of correlations today to the variation of IP growth rates one year ago. Therefore, we only look back for three months. For such short lag period, the difference between flat weights and uneven weights becomes negligible, so we select flat weights for computational simplicity Results Before discussing numerical results, let us examine some figures to have a general impression of these models. Figure 1.2 plots the quarterly realized correlations of the stock and bond, and the correlations exhibit strong temporal variations. These variations support the applications of dynamic copulas, because no static copula can produce such volatile patterns. Figure 1.3 further plots the quarterly realized correlations along with the implied correlations of GAS model. It shows that the implied correlations closely follow the realized ones while the realized ones have wilder fluctuations. In Figure 1.4, we reduce the sampling window of realized correlations to one month and plot the monthly realized correlations with the implied correlations of GAS model. Now the implied correlations seem to be a long term component of the monthly realized correlation. These two figures convey that GAS model is highly persistent and similar patterns can also be observed with GAS-based models through Figure 1.5 to Figure 1.7. All the estimation results are presented in Table 1 to 5. The last rows of these tables contain the likelihoods of the six models. By comparing these likelihoods, we make a number of observations on the model performances. First, GAS-based models have 14

26 higher likelihoods than PROD model driven by cross products. Moreover, GAS-DCC and GAS-ADD models are better than GAS model, and GAS-DCC-MIDAS model is better than GAS-DCC model in terms of likelihood. Since GAS and GAS-ADD models are nested, we can apply likelihood ratio tests to compare these two models. It can be shown that the difference is statistically significant. Similar arguments apply to the pairs of GAS/GAS-DCC and GAS-DCC/GAS-DCC-MIDAS. By contrast, GAS-MIDAS model does not offer much more than the GAS model. This may suggest that it is better to weight long term variations by macro variables than short term variations. Now let us turn our attention to interpreting parameter estimates. The first rows of these tables report the estimates of α 1, reflecting the influence of daily innovations on dependence parameter. For all the six models, this parameter is significant and has a positive sign as expected. Furthermore, the estimates of GAS-DCC, GAS-DCC-MIDAS and GAS-ADD models have much bigger values than the one of GAS model. This may imply that the former models can offer more volatilities than the latter one. The second rows of these tables refer to the estimates of α 2. This parameter measures the influence of long term (weekly) innovations on the dependence parameter of GAS- DCC and GAS-DCC-MIDAS models. The estimates of both models are significant and have negative signs. These negative signs seem to remove part of the daily innovations accumulated in the past week, and make GAS-DCC and GAS-DCC-MIDAS models more volatile than other models. Figure 1.6 plots the implied correlations of GAS and GAS- DCC models. GAS-DCC model has a thicker curve than GAS model, and this also conveys that the former model has more volatility than the latter. Besides α 1 and α 2, θ 1 and θ 2 also affect how innovations integrate into the dependence parameters. For most of the five industries, these two parameters are significant. In GAS- DCC-MIDAS model, θ 1 are positive and θ 2 are negative for all five industries. This means that (θ 1 X t 1 Xt<0 + θ 2 X t 1 Xt 0) < 0 and τ t < 1 always. When the absolute value of X t is 15

27 higher, θ 1 X t 1 Xt<0 + θ 2 X t 1 Xt 0 is smaller and α 2 /τ t is larger. A large α 2 /τ t removes more weekly innovations from the dependence parameter, providing more fluctuations to the model. Figure 1.5 also demonstrates that the implied correlations of GAS-DCC- MIDAS model are more volatile than the ones of GAS-DCC model. That s probably why GAS-DCC-MIDAS model has a better performance than the GAS-DCC model. θ 1 and θ 2 also appear in GAS-MIDAS model. Since GAS-MIDAS model is not much different from GAS model in terms of likelihoods, we skip interpreting these two parameters for GAS-MIDAS model. While all the parameters mentioned above measure the loading of innovation terms, β measures the influence of lagged terms. For all the six models except GAS-ADD model, β are fairly close to one, and GAS based models have higher values than PROD model. These estimates are consistent with the results in other papers also using GAS models, say Patton (2012). GAS-ADD model is a special case, since its β is lower than 80 percent. But if we add up β and ω, we find that the sums are also close to one in all industries. So GAS-ADD model transfers part of the weight on the lagged implied correlation to the intercept that is the long term component in the model. This model still yields highly persistent implied correlations. ν measures the tail thickness of the t copula. The larger the ν value is, the thinner the tail is. Compare the ν value of GAS-DCC/GAS-DCC-MIDAS/GAS-ADD model with that of PROD/GAS model, we find that the former model has a higher value than the latter, i.e.: the former model has a thinner tail than the latter, and less extreme events happen with the former models. This comparison conveys that we reduce the tail thickness by providing more accurate estimates of the dependence parameter. Some further comparisons can be drawn between GAS-ADD and GAS-DCC models. Comparing the coefficients and likelihoods of these two models, we find that all these outputs are rather similar: the values of likelihoods are almost identical; the estimates 16

28 of α 1 and ν are alike; the loadings on lagged terms are both close to one. This similarity can be further confirmed by examining the implied correlations of these two models in Figure 1.7, and the two curves of implied correlations are almost the same. These coincidences reveal the interconnection between GAS-DCC and GAS-ADD models and are possibly caused by the high persistence of GAS models. Since β is approximately one, it influences the model equivalently by adding a time varying intercept or a weekly innovation. 1.6 Conclusions In this chapter, we propose a novel class of dynamic copula models and extract long and short term components from the dependence parameter. An extra component adds flexibility to the model, and most of the empirical applications show improved prediction. Moreover, we introduce the MIDAS framework to dynamic copulas and combine daily returns with monthly updated macro variables. Specifically, we use the macro variable as a proxy for the market condition, and allow that the market condition affects how dependence parameter reacts to innovations. We find that introducing macro variables adds more volatility to the dependence parameter, and therefore improved the model performance consistently through the empirical applications. 17

29 Table 1.1: Estimation Results for Consumer Goods PROD GAS GAS-DCC GAS-MIDAS GAS-DCC-MIDAS GAS-ADD α (5.40) (5.58) (19.59) (4.95) (5.37) (6.53) α (-49.40) - (-2.56) - β (293.82) (939.56) ( ) (366.50) ( ) (22.44) θ (-11.15) (2.92) - θ (2.73) (-3.61) - ν (9.55) (8.54) (85.68) (431.68) (8.42) (8.96) ω (-2.97) (0.03) (0.04) (0.13) (0.87) (3.86) logl Notes: This table reports the estimates for the dynamic copula models with 10-year Treasury bond and the stock portfolio of the consumer goods industry. T-statistics are in parentheses. Table 1.2: Estimation Results for Manufacturing PROD GAS GAS-DCC GAS-MIDAS GAS-DCC-MIDAS GAS-ADD α (6.01) (4.58) (10.97) (33.95) (6.12) (7.68) α (-7.83) - (-2.96) - β (350.64) (569.77) ( ) (752.08) ( ) (21.10) θ (-67.46) (6.48) - θ (25.89) (-4.56) - ν (8.32) (26.53) (524.23) ( ) (7.93) (123.70) ω (-3.64) (-0.23) (-0.22) (-0.07) (1.01) (3.90) logl Notes: This table reports the estimates for the dynamic copula models with 10-year Treasury bond and the stock portfolio of the manufacturing industry. T-statistics are in parentheses. 18

30 Table 1.3: Estimation Results for Health PROD GAS GAS-DCC GAS-MIDAS GAS-DCC-MIDAS GAS-ADD α (5.86) (4.58) (10.97) (33.95) (6.12) (7.68) α (-7.83) - (-2.96) - β (332.80) ( ) ( ) ( ) ( ) (7.80) θ (-1.79) (2.06) - θ (0.58) (-0.15) - ν (9.05) (8.64) (8.43) (8.39) (8.30) (0.42) ω (-3.18) (0.30) (0.35) (0.18) (-0.18) (0.90) logl Notes: This table reports the estimates for the dynamic copula models with 10-year Treasury bond and the stock portfolio of the health industry. T-statistics are in parentheses. Table 1.4: Estimation Results for HiTec PROD GAS GAS-DCC GAS-MIDAS GAS-DCC-MIDAS GAS-ADD α (6.04) (6.29) (4.41) (5.45) (4.44) (4.16) α (-2.75) - (-2.24) - β (329.57) ( ) ( ) ( ) ( ) (17.04) θ (-1.53) (4.23) - θ (0.19) (-0.99) - ν (8.27) (123.55) (6.88) (7.73) (7.44) (18.56) ω (-2.95) (0.13) (0.12) (-0.59) (0.14) (2.40) logl Notes: This table reports the estimates for the dynamic copula models with 10-year Treasury bond and the stock portfolio of the HiTec industry. T-statistics are in parentheses. 19

31 Table 1.5: Estimation Results for Other PROD GAS GAS-DCC GAS-MIDAS GAS-DCC-MIDAS GAS-ADD α (6.45) (3.96) (7.09) (7.05) (6.48) (6.99) α (-4.41) - (-3.33) - β (332.55) (630.02) ( ) ( ) ( ) (27.79) θ (-2.18) (5.54) - θ (0.90) (-5.05) - ν (10.58) (383.07) (8.85) (9.52) (8.88) (8.94) ω (-4.51) (-0.39) (-0.25) (-0.64) (-1.12) (3.31) logl Notes: This table reports the estimates for the dynamic copula models with 10-year Treasury bond and the stock portfolio of the other industry. T-statistics are in parentheses. 20

32 Figure 1.1: Industrial Production and Realized Correlations Realized Correlation Date Growth Rate Date Notes: The upper panel shows the quarterly realized correlations of 10-year Treasury bond and the stock portfolio of manufacturing industry. The lower panel presents the monthly growth rates of industrial production in U.S. 21

33 Figure 1.2: The Quarterly Realized Correlations of the Stock and Bond Realized Correlation Date Notes: This picture reports the quarterly realized correlations of 10-year Treasury bond and the stock portfolio of manufacturing industry. The realized correlations are calculated on a rolling basis. Figure 1.3: The Quarterly Realized Correlation and Implied Correlations Correlation Implied Realized Date Notes: This picture reports the correlations of 10-year Treasury bond and the stock portfolio of manufacturing industry. The dark line shows the implied correlations of GAS model. The light line represents the quarterly realized correlations calculated on a rolling basis. 22

34 Figure 1.4: The Monthly Realized Correlations and the Implied Correlations Correlation Implied Realized Date Notes: This picture reports the correlations of 10-year Treasury bond and the stock portfolio of manufacturing industry. The dark line shows the implied correlations of GAS model. The light line represents the monthly realized correlations calculated on a rolling basis. Figure 1.5: The Implied Correlations of GAS-DCC and GAS-DCC-MIDAS Models Correlation GAS DCC GAS DCC MIDAS Date Notes: This picture reports the correlations of 10-year Treasury bond and the stock portfolio of manufacturing industry. The dark line shows the implied correlations of GAS-DCC model. The light line represents the implied correlations of GAS-DCC- MIDAS model. 23

35 Figure 1.6: The Implied Correlations of GAS and GAS-DCC Models The Implied Correlations of GAS Model Implied Correlation Date The Implied Correlations of GAS DCC Model Implied Correlation Date Notes: This picture reports the correlations of 10-year Treasury bond and the stock portfolio of manufacturing industry. The upper panel shows the implied correlations of GAS model. The lower panel presents the implied correlations of GAS-DCC model. 24

36 Figure 1.7: The Implied Correlations of GAS-DCC and GAS-ADD Model The Implied Correlations of GAS DCC Model Implied Correlation Date The Implied Correlations of GAS ADD Model Implied Correlation Date Notes: This picture reports the correlations of 10-year Treasury bond and the stock portfolio of manufacturing industry. The upper panel shows the implied correlations of GAS-DCC model. The lower panel presents the implied correlations of GAS-ADD model. 25

37 CHAPTER 2: FRAILTY MODELS FOR COMMERCIAL MORTGAGES 2.1 Introduction Credit risk affects virtually all aspects of financial activities. It is an important factor in pricing financial products and has profound influence on risk management. Moreover, policy makers and regulators pay special attention to credit risk when they design economic policies and regulatory frameworks. There is an extensive literature on the measurement and management of credit risk. Researchers have used various approaches to model the credit risk of corporate debts, mortgages and derivatives. In general, these models can be divided into two categories structural models and reduced-form models. This article presents a reduced-form model that contains frailty factors to predict mortgage default. Reduced-form models rely on two sources of information to explain credit risk. One source is the financial information of borrowers, which is employed to track the idiosyncratic part of the credit risk. The other source comes from macro variables, which approximates the systematic part. Failure to explain all the systematic risk would introduce biases when estimating risk measures. Das et al. (2007) while studying corporate bond defaults provided evidence that macro variables alone were not enough to explain all the systematic risk. They further demonstrated that the lack of explanatory variables underestimates value-at-risk. To produce unbiased estimates, they proposed frailty models for corporate bond credit risk, to account for the unexplained part of systematic risk. Frailty models can be classified into two types - using a characterization put forward 26

38 by Cox et al. (1981): parameter-driven models and observation-driven models. Both types of models have been used to predict credit risk. For example, parameter-driven models have been used to track the credit risk of corporate debts in Duffie et al. (2009) and Koopman and Lucas (2008), and to forecast mortgage default in Kau, Keenan and Li (2011). Meanwhile, using observation-driven models, Creal, Koopman and Lucas (2013a) and Creal et al. (2014) investigated corporate defaults. The predictability of latent factors is a main feature of observation-driven models. This feature indicates that current period frailty factors can be computed using only past information. In contrast, the computation of frailty factors in parameter-driven models not only requires past information but also future and current information. Inference of parameter-driven models therefore generally requires simulation, which is time consuming with large data sets. The estimation of observation-driven models is in comparison rather straightforward. In this chapter, we develop a novel framework to model systematic risk of mortgages. Specifically, we match default rates in multiple dimensions by extending the generalized autoregressive score (GAS) models proposed in Creal, Koopman and Lucas (2013a). Our data consists of commercial mortgages in the U.S. multifamily market from 1997 to We construct a series of models and employ multiple tests to demonstrate the advantages of our framework. To the best of our knowledge, this is the first attempt that uses observation-driven models to predict mortgage defaults. We show that the new class of models we propose has better tractability compared with parameter-driven models. For instance, although our dataset has more than two million records, and our most complex model incorporates up to 15 frailty factors, the estimation process only takes two minutes using a standard desktop computer. Compare this with for example Kau, Keenan and Smurov (2006) who employed parameter-driven models to predict mortgage defaults using a small data set. Their method requires simulations and is very time consuming and therefore practically 27